feat(vad): add Silero VAD adapter

[wuhongsheng]

Summary

• add vad_model="silero-vad" backed by the official silero-vad Python package
• preserve the existing AutoModel VAD → ASR pipeline contract by returning millisecond segments
• add optional dependency, documentation, and focused tests

Validation

• python -m pytest -q tests/test_silero_vad_adapter.py
• README example tested with cloth_20220629_clean_01.wav (non-empty transcript and 329 timestamps)

[LauraGPT]

Maintainer validation completed on e8ea59a02c9362ae78e6730ca574862f5aedd6c7. I pushed two focused fixes before moving this out of draft:

• keep TorchScript input tensors on the adapter/model device so CUDA ASR + Silero VAD does not mix a CUDA model with CPU audio
• keep ONNX input tensors on CPU because the official wrapper calls Tensor.numpy()
• reject negative and sub-millisecond max segment values that would otherwise produce a zero-width split loop
• add the optional Silero path to the main English README

Validation:

• python -m pytest -q tests/test_silero_vad_adapter.py -> 7 passed with silero-vad==6.2.1
• real 70.47 s WAV through the TorchScript alias -> 42 valid segments
• the same WAV through silero_onnx=True with onnxruntime==1.28.0 -> 42 valid segments
• Black check, py_compile, and git diff --check passed
• repository contract check passed

[LauraGPT]

Reviewed the final maintainer-tested head. TorchScript and ONNX device behavior, segment bounds, optional dependency path, and documentation are covered.